Switching Infinite and Finite Sums (Markov Chains and Mixing Times Theorem 20.3)

117 Views Asked by At

I was working through Theorem 20.3 Part (i) from Markov Chains and Mixing Times (Levin, Peres, Wilmer) and I am confused about the justification of an inequality. From the book (p. 283):

Proof of Theorem 20.3. (i), Step 1. Recall that $N_t$ is the Poisson random variable indicating the number of transitions in the continuous-time chain. We first prove $$ \left\|H_t(x, \cdot)-\pi\right\|_{\mathrm{TV}} \leq \mathbf{P}\left\{N_t<k\right\}+\left\|P^k(x, \cdot)-\pi\right\|_{\mathrm{TV}} . $$ Conditioning on the value of $N_t$ and applying the triangle inequality give $$ \left\|H_t(x, \cdot)-\pi\right\|_{\mathrm{TV}} \leq \sum_{j \geq 0} \mathbf{P}\left\{N_t=j\right\}\left\|P^j(x, \cdot)-\pi\right\|_{\mathrm{TV}} $$

My question: I feel like I am missing something simple, but why can one get the inequality from applying the triangle inequality alone? My thought was that, because you have an infinite sum, you would need Fubini-Tonelli or something like this to exchange the sums? The following is my attempt at a proof:

By definition (p.281), the Heat Kernel with transition matrix P is

$$ \begin{aligned} H_t(x, y) & =\sum_{k=0}^{\infty} \frac{e^{-r t}(r t)^k}{k !} P^k(x, y) \end{aligned} $$

Thus, using $P^k_x$ to denote the distribution at time $k$ starting from $x$, and $(H^t_x$ is analogous.)

$$ \begin{align} \begin{split} \|H^t_x - \pi\|_{TV}& =\left\| \left(\sum_{k=0}^{\infty} \frac{e^{-r t}(r t)^k}{k !} P_x^k\right) - \pi\right\|_{TV}\\ &\leq\left|\sum\limits_{\omega \in A}\left(\sum_{k=0}^{\infty} \frac{e^{-r t}(r t)^k}{k !} P_x^k(\omega)\right) - \pi(\omega)\right|\\ \end{split} \end{align} $$

Using A to denote the set which maximizes the absolute value of the difference between $P_x^k$ and $\pi$ , then this line follows from the definition of TVD.

Then by Fubini-Tonelli

$$\left|\sum_{k=0}^{\infty} \frac{e^{-r t}(r t)^k}{k !} \sum\limits_{\omega \in A} P_x^k(\omega) - \pi(\omega)\right| $$

Then by Triangle inequality

$$ \begin{align} \begin{split} &\leq\sum_{k=0}^{\infty}\left| \frac{e^{-r t}(r t)^k}{k !} \sum\limits_{\omega \in A} P_x^k(\omega) - \pi(\omega)\right|\\ &=\sum_{k=0}^{\infty}\left| \frac{e^{-r t}(r t)^k}{k !}\right| \left|\sum\limits_{\omega \in A} P_x^k(\omega) - \pi(\omega)\right|\\ &\leq\sum_{k=0}^{\infty}\frac{e^{-r t}(r t)^k}{k !} \|P_x^k-\pi \|_{TV}\\ &=\sum_{k=0}^{\infty}P(N(rt)=k) \|P_x^k-\pi \|_{TV}\\ \end{split} \end{align} $$

where N(rt) is a poisson of rate $rt$. And this is the same as the given inequality. Please do let me know if I've made a mistake in my proof. If not though, I'm not sure how you would get here without using something to interchange the sum from TVD and the infinite sum.

1

There are 1 best solutions below

5
On BEST ANSWER

Based on the comments your definition of $A$ wanted to be the maximizer event in the definition of $\|H_x^t-\pi\|_{TV}$, which can be written as $A = \{\omega\in\mathcal{X} : H_x^t(\omega) \ge \pi(\omega)\}$ by Remark 4.3 in Markov Chains and Mixing Times (Levin, Peres, Wilmer). Also, based on your observation, we have $$ \left|\sum_{\omega \in A}P_x^k(\omega)-\pi(\omega)\right| \le \|P_x^k - \pi\|_{TV} $$ by the definition of $\|\cdot\|_{TV}$ for any $A \subseteq \mathcal{X}$.

Hence it remains to prove that you can apply Fubini's theorem, which is true if: $$ \sum_{\omega\in A}\sum_{k=0}^\infty \frac{e^{-rt}(rt)^k}{k!} \left|P_x^k(\omega)-\pi(\omega)\right| < \infty . $$ But this holds by using $\sum_{\omega\in A}H_x^t(\omega) \le 1$ and $\sum_{\omega \in A}\pi(\omega) \le 1$ as they are distributions over the state space $\mathcal{X}$ and $A \subseteq \mathcal{X}$, and then $$ \sum_{\omega\in A}\sum_{k=0}^\infty \frac{e^{-rt}(rt)^k}{k!} \left|P_x^k(\omega)-\pi(\omega)\right| \le \sum_{\omega\in A}\sum_{k=0}^\infty \frac{e^{-rt}(rt)^k}{k!} \Big(P_x^k(\omega) + \pi(\omega)\Big) = \sum_{\omega\in A} H_x^t(\omega) + \pi(\omega) \le 2 . $$ So it seems that your proof is correct after all. Nice work! :)